University College London
MPhil / PhD · Computer Science
I’m a PhD student in Computer Science at University College London, advised by Professor Lorenzo Cavallaro in the Systems Security Research Lab (S2Lab). Previously, I completed my MEng degree in Computing (Artificial Intelligence and Machine Learning) at Imperial College London.
My research interests center on the systems around agents: context management, memory, tool use, routing, and verification. I study how these components can help agents reason over the right evidence, use computation efficiently, and improve from execution experience. I evaluate this work across diverse agent benchmarks and real-world domains, including program analysis and malware understanding.
Education
Imperial College London
MEng · Computing
Internships
Microsoft
Research Intern
Developing quality- and cost-efficient agent harness optimization across coding and general-purpose agent settings.
BAFTA Media Technology
Software Engineering Intern
Built BAFTA View experiences for Fire TV and the web, with a focus on accessible cross-platform interaction.
Shanghai Ximalaya Technology
Software Engineering in Test Intern
Designed automated API testing infrastructure and led smoke and regression testing for production services.
ResearchContext management Grounding Efficiency Self-improvement
My research studies how agent systems turn interaction into reliable improvement through four connected capabilities.
Selecting, retaining, and routing the information agents need.
Connecting reasoning and actions to verifiable evidence.
Improving capability while controlling computation and cost.
Turning execution experience into auditable system updates.
Publications
ParetoSaddler: Pareto-Guided Optimization of Agent Harnesses for Quality and Cost
ParetoSaddler jointly optimizes agent quality and cost by separating intermediate candidate promotion from final deployment selection, using Bayesian bootstrap Pareto support to retain credible improvements across ClawMark and GAIA2.
arXiv preprint arXiv:2611.02079
Page Paper Code
BibTeX
@article{qian2026paretosaddler,
title = {{ParetoSaddler}: Pareto-Guided Optimization of Agent Harnesses for Quality and Cost},
author = {Qian, Xingzhi and Park, Sungho and Tan, Rongyuan and Yu, Jianbo and Zhang, Jue and Cavallaro, Lorenzo and Xia, Menglin and Chen, Yanzhi and Xu, Xinnuo and Kang, Yu and Zheng, Xinran and Somashekar, Gagan and Zhang, Xuchao and Bansal, Chetan and Lin, Qingwei and Rühle, Victor},
journal = {arXiv preprint arXiv:2611.02079},
year = {2026}
}Malaika: Understanding Malware through Tri-Grounded Agentic Reasoning
Malaika structures multi-agent malware understanding around domain, semantics, and knowledge grounding, combining adaptive evidence localization with explicit review to produce precise, auditable conclusions at practical cost.
arXiv preprint arXiv:2607.09179
Paper
BibTeX
@article{qian2026malaika,
title = {Malaika: Understanding Malware through Tri-Grounded Agentic Reasoning},
author = {Qian, Xingzhi and Zheng, Xinran and He, Yiling and Cavallaro, Lorenzo},
journal = {arXiv preprint arXiv:2607.09179},
year = {2026}
}Is ‘Knowing It’s Malicious’ Enough? Evaluating LLMs for Fine-Grained Malware Behavior Auditing
A reproducible framework that asks whether LLMs can prioritize functions, attribute evidence, synthesize behavior, and discriminate samples — with analyst-verifiable outputs.
35th ACM SIGSOFT International Symposium on Software Testing and Analysis
Paper Code
BibTeX
@inproceedings{zheng2026knowing,
title = {Is {Knowing It's Malicious} Enough? Evaluating LLMs for Fine-Grained Malware Behavior Auditing},
author = {Zheng, Xinran and Qian, Xingzhi and He, Yiling and Yang, Shuo and Cavallaro, Lorenzo},
booktitle = {Proceedings of the 35th ACM SIGSOFT International Symposium on Software Testing and Analysis},
year = {2026}
}On Benchmarking Code LLMs for Android Malware Analysis
Cama evaluates how code LLMs interpret decompiled Android code through consistency, fidelity, and semantic relevance across 118 malware samples and 7.5M+ functions.
34th ACM SIGSOFT International Symposium on Software Testing and Analysis
Paper
BibTeX
@inproceedings{he2025benchmarking,
title = {On Benchmarking Code LLMs for Android Malware Analysis},
author = {He, Yiling and She, Hongyu and Qian, Xingzhi and Zheng, Xinran and Chen, Zhuo and Qin, Zhan and Cavallaro, Lorenzo},
booktitle = {Companion Proceedings of the 34th ACM SIGSOFT International Symposium on Software Testing and Analysis},
year = {2025}
}LAMD: Context-driven Android Malware Detection and Classification with LLMs
LAMD extracts security-critical context and builds program structures for tier-wise reasoning, pairing detection with explanations and factual-consistency checks.
46th IEEE Symposium on Security and Privacy Workshops
Paper
BibTeX
@inproceedings{qian2025lamd,
title = {{LAMD}: Context-driven Android Malware Detection and Classification with {LLMs}},
author = {Qian, Xingzhi and Zheng, Xinran and He, Yiling and Yang, Shuo and Cavallaro, Lorenzo},
booktitle = {2025 IEEE Symposium on Security and Privacy Workshops},
year = {2025}
}